munish0838/crop-burn-detection-labeled

Crop Burn Detection — Labeled Sentinel-2 (India, 2025)

Paired RGB + SWIR Sentinel-2 satellite image tiles across agricultural districts of northern India, with per-tile burn annotations. Covers the paddy (Oct–Nov 2025) and wheat (Mar–May 2025) stubble burning seasons across 68 districts. Built to train and benchmark vision models for real-time crop residue burn detection.

The raw (unlabeled) version is available as munish0838/crop-burn-detection-raw.

---

Why We Built This

Every October and November, farmers across Punjab, Haryana, Uttar Pradesh, Rajasthan, and Madhya Pradesh burn millions of tonnes of paddy stubble left after harvest. It is the fastest and cheapest way to clear fields before the next crop — but the smoke blankets northern India in a toxic haze that turns Delhi's AQI past 500, sends millions to hospitals, and contributes meaningfully to South Asia's greenhouse gas load.

The same cycle repeats in April–May for wheat.

Satellite imagery already captures this burning as it happens. What's missing is a model that can detect, quantify, and track burns at scale and in near real-time — one that is lightweight enough to run on-board next-generation Earth observation satellites, eliminating the download bottleneck entirely.

This dataset is the foundation for that model. We assembled 5 km × 5 km Sentinel-2 tiles centered on known agricultural fire hotspots across 68 districts, sampled every 5–10 days across both burning seasons. Every tile includes both a natural-color RGB composite and a SWIR false-color composite — the combination that remote sensing practitioners use to distinguish active burns and recent scars from ordinary harvested fields.

Each tile is labeled with burn presence, severity, affected fraction, freshness, vegetation phase, and quality flags — derived from pixel-level spectral analysis using the NBR-like index (NIR − SWIR2.2) / (NIR + SWIR2.2) computed directly from the SWIR composite bands.

---

Dataset at a Glance

	Train	Test
Pairs	1,098	267
Districts	55	13
States	Punjab, Haryana, UP, MP, Rajasthan	Punjab, UP, MP, Rajasthan
Seasons	Paddy 2025 + Wheat 2025	Paddy 2025 + Wheat 2025

• Total labeled pairs: 1,365
• Tile size: 5 km × 5 km
• Spatial resolution: ~10 m/pixel (Sentinel-2)
• Split strategy: District-level — no district appears in both train and test
• Country: India

---

Composites Explained

Each date and location has two images:

rgb_image — Natural Color (B4-B3-B2)

Standard visible composite. Healthy crops appear green, harvested stubble appears tan/brown, burn scars appear dark gray to black, water appears dark blue.

swir_image — SWIR False Color (B12-B8-B4)

Shortwave infrared composite, the key band for burn detection.

What you see	What it means
Bright green	Healthy standing crop (high NIR)
Brownish / pinkish-brown	Harvested stubble or bare soil
Dark brown / charcoal black	Burn scars — the darker and sharper, the fresher
White / bright cyan	Cloud cover
Solid black	Water or missing data

---

Labeling Methodology

Labels are derived from pixel-level spectral analysis of the SWIR composite:

• NBR-like index: (G − R) / (G + R) computed from SWIR PNG pixel values, where G = NIR channel, R = SWIR2.2 channel
• Burn detection: Pixels with negative or low NBR, appropriate brightness profiles, and RGB confirmation are classified as burns
• Three burn stages detected: Fresh burns (charcoal-black, very dark), recent burns (brownish-red, negative NBR), old burns (faded, slightly negative NBR)
• Quality masking: Cloud pixels (white/bright-cyan in SWIR) and missing data (near-black in both images) are masked before computing statistics
• Thresholds: Burn detected when ≥0.5% of valid (non-cloud, non-missing) cropland pixels show burn signature

---

Schema

{
  # Image pair
  "image_id":               str,   # e.g. "train_paddy_PUNJAB_AMRITSAR_2025-10-01"
  "rgb_image":              Image, # PIL image, natural color (B4-B3-B2)
  "swir_image":             Image, # PIL image, SWIR false color (B12-B8-B4)
  # Metadata
  "split":                  str,   # "train" or "test"
  "season":                 str,   # "paddy" or "wheat"
  "date":                   str,   # "YYYY-MM-DD"
  "district":               str,   # e.g. "AMRITSAR"
  "state":                  str,   # e.g. "PUNJAB"
  "country":                str,   # "India"
  "latitude":               float, # tile center latitude
  "longitude":              float, # tile center longitude
  "tile_size_km":           float, # always 5.0
  "fires_district_2025":    int,   # CREAMS fire count for this district in 2025
  # Labels
  "burn_detected":          bool,  # True if burn scars detected
  "burn_severity":          str,   # "none" | "low" | "moderate" | "high"
  "burn_fraction_estimate": float, # fraction of visible cropland burned (0–1, rounded to 0.05)
  "burn_freshness":         str,   # "none" | "fresh" | "recent" | "old"
  "active_smoke_visible":   bool,  # True if active smoke plume visible
  "vegetation_phase":       str,   # "pre_harvest" | "harvest_in_progress" | "post_harvest" | "fallow"
  "image_quality_limited":  bool,  # True if >30% cloud/missing data
  "notes":                  str,   # Human-readable description of the scene
  "annotation_json":        str,   # Complete annotation as a JSON string (for VLM/generative training)
}

Label Definitions

Field	Values	Description
burn_severity	none / low / moderate / high	Low: ≤10% cropland burned; Moderate: 10–40%; High: >40%
burn_freshness	none / fresh / recent / old	Fresh: charcoal-black, very dark; Recent: brownish-red, negative NBR; Old: faded, slightly negative NBR
vegetation_phase	pre_harvest / harvest_in_progress / post_harvest / fallow	Based on observed NIR vegetation signal in SWIR composite
burn_fraction_estimate	0.0 – 1.0	Proportion of visible cropland showing burn signature, rounded to nearest 0.05

---

Coverage

Paddy Season (October – November 2025)

12 observation dates per district: 2025-10-01, 2025-10-10, 2025-10-15, 2025-10-20, 2025-10-25, 2025-10-30, 2025-11-04, 2025-11-09, 2025-11-14, 2025-11-19, 2025-11-24, 2025-11-29

Wheat Season (March – May 2025)

14 observation dates per district: 2025-03-25, 2025-04-01, 2025-04-06, 2025-04-11, 2025-04-16, 2025-04-21, 2025-04-26, 2025-05-01, 2025-05-06, 2025-05-11, 2025-05-16, 2025-05-21, 2025-05-26, 2025-05-31

Districts by State

State	Train districts	Test districts
Punjab	11	3
Uttar Pradesh	18	7
Madhya Pradesh	20	2
Rajasthan	6	1
Haryana	1	0

---

How to Use

from datasets import load_dataset

ds = load_dataset("munish0838/crop-burn-detection-labeled")

# Access a sample
sample = ds["train"][0]
sample["rgb_image"].show()   # PIL Image
sample["swir_image"].show()  # PIL Image
print(sample["date"], sample["district"], sample["burn_detected"], sample["burn_severity"])

# Filter to burn-positive samples only
burns = ds["train"].filter(lambda x: x["burn_detected"])

# Filter to Punjab paddy season
punjab_paddy = ds["train"].filter(
    lambda x: x["state"] == "PUNJAB" and x["season"] == "paddy"
)

# Filter to high-quality images (not cloud-limited)
clear_sky = ds["train"].filter(lambda x: not x["image_quality_limited"])

# Stream without downloading everything
ds_stream = load_dataset("munish0838/crop-burn-detection-labeled", streaming=True)

---

Intended Uses

• Burn detection / classification — binary or multi-class burn presence
• Burn severity estimation — fraction of cropland affected
• Temporal change detection — tracking burn progression across dates
• VLM fine-tuning — vision-language models for on-board satellite inference
• Agricultural monitoring — harvest timing and crop phase detection
• Air quality forecasting — burn activity as upstream input to pollution models

---

Limitations

• Tiles with heavy cloud cover (>30%) are included but flagged via image_quality_limited
• Some tiles have partial no-data regions (edge of satellite swath)
• Coverage is limited to districts with historically high fire activity (≥100 CREAMS fires in 2025) — low-fire regions are underrepresented
• All tiles are from 2025; year-to-year variability is not captured in this release
• Labels are derived from pixel-level spectral analysis, not manual visual inspection of each tile

---

Data Source

Imagery sourced from Sentinel-2 (ESA Copernicus programme), 5-day composites. Fire reference data from CREAMS (Crop Residue Burning Emission and Monitoring System). Tile selection uses fire-centroid sampling to center tiles on historically active burn areas.

---

License

Creative Commons Attribution 4.0 (CC BY 4.0) — free to use for research and commercial purposes with attribution.

Sentinel-2 imagery is provided under the Copernicus Data Policy.